Shin Sangyun ; Zhou Kaichen ; Vankadari Madhu ; Markham Andrew ; Trigoni Niki

Abstract

Coarse-to-fine 3D instance segmentation methods show weak performancescompared to recent Grouping-based, Kernel-based and Transformer-based methods.We argue that this is due to two limitations: 1) Instance size overestimationby axis-aligned bounding box(AABB) 2) False negative error accumulation frominaccurate box to the refinement phase. In this work, we introduce SphericalMask, a novel coarse-to-fine approach based on spherical representation,overcoming those two limitations with several benefits. Specifically, ourcoarse detection estimates each instance with a 3D polygon using a center andradial distance predictions, which avoids excessive size estimation of AABB. Tocut the error propagation in the existing coarse-to-fine approaches, wevirtually migrate points based on the polygon, allowing all foreground points,including false negatives, to be refined. During inference, the proposal andpoint migration modules run in parallel and are assembled to form binary masksof instances. We also introduce two margin-based losses for the point migrationto enforce corrections for the false positives/negatives and cohesion offoreground points, significantly improving the performance. Experimentalresults from three datasets, such as ScanNetV2, S3DIS, and STPLS3D, show thatour proposed method outperforms existing works, demonstrating the effectivenessof the new instance representation with spherical coordinates. The code isavailable at: https://github.com/yunshin/SphericalMask